Temperature

Temperature is a Sampling parameter that sharpens or smooths a language model's probability distribution — low values push the model toward the most likely tokens, while higher ones flatten the distribution and admit more surprising completions. Values near 0 are typical for deterministic tasks like code or data extraction, while 0.7-1.0 is the usual range for creative writing. Most APIs default to somewhere between 0.7 and 1.0, but even 'temperature 0' is rarely fully deterministic — hardware and Sampling implementations leave tiny variations. Combined with Top-K or Top-P, temperature gives you dramatic control over the character of the output.